| Abstract [eng] |
Conventional autopilot systems fail to navigate safely if the GPS signal is lost, jammed or unavailable. Localization, the process of pose estimation relative to a known environment, may solve the problem of navigation in GPS-Denied environment. Visual odometry, Simultaneous Localization and Mapping can be used to process aerial imagery from a downward-facing camera onboard UAV may be used to solve the pose estimation problem. But these methods are prone to errors over long-distance flights (>1 km), while map-based technique can provide additional accuracy. A map-based technique is proposed in this dissertation called Discriminatory Pearson Correlation based Particle Filter Localization (abbr. DCP-PFL). The algorithm is developed using KLD sampling based Particle filter and Pearson correlation. The experimental results obtained in this dissertation identified that Pearson Correlation is the most suitable image similarity metric to match aerial images to a map and KLD sampling technique provides the results 3x times faster with similar results compared to other techniques. The main contribution of the work is the proposed image similarity to probability conversion functions. The conversion functions are parametrized to achieve a trade-off between localization accuracy and robustness to inaccuracies in the map. DCP-PFL algorithm is compared against state-of-the-art algorithms SVO and ORB-SLAM. The proposed algorithm reduced localization error by a factor of 2x and reduced the error drift slope by 11x times compared to the SVO algorithm. Comparing against ORB-SLAM, the proposed DCP-PFL algorithm achieved 2.6 times better accuracy. |
